Annotation and classification of Political advertisements

Political advertising has changed drastically over the last several decades with video advertising becoming a major force in all outlets from the traditional TV medium to online media. In this work, we attempt to automatically classify the political advertisements along various dimensions such as purpose, content and emotion. First, we use a crowd-sourcing method to annotate the political advertisements in terms of how viewers perceive them along the above mentioned aspects. Then, we use audio-based features and machine learning algorithms for automatic classification tasks. In particular, we deploy speech-related features along with support vector machine (SVM) and music-related features along with k-nearest neighbor (KNN). The analysis of crowd-sourced annotations shows that the same advertisements are often used to serve multiple purpose and that certain content categories such as speech clips from the candidate and other public figures are more prevalent. The experimental results using speech/audio features on advertisements aired during the U.S. presidential campaign of 2012 show promising classification performance.